Definition and measurement of tree cover: A comparative analysis of field-, lidar- and landsat-based tree cover estimations in the Sierra national forests, USA

Abstract The accuracy of mapping forest extent from satellite imagery largely depends on both semantic definition and estimation quality of tree cover at the pixel level. In this study, we conducted a comparative analysis of different tree cover data sets, derived from field sample, terrestrial and airborne lidar scans, and Landsat imagery, to investigate factors affecting tree cover estimation in a western mountainous conifer forest in the United States. We found that satellite-based tree cover maps, even those derived from the same sensor, tended to have lower agreements over sparsely vegetated areas, due to both definitional discrepancies and estimation errors in establishing tree cover. When compared with tree cover measurements from ground-based hemispherical photography and terrestrial laser scanner (TLS), both airborne waveform lidar and discrete return lidar (DRL) provided consistently more accurate tree cover estimates (  r 2 ranging from 0.61 to 0.91, RMSE ≈ 15% and bias r 2 ranging from 0.64 to 0.78, RMSE ≈ 20% and bias varying from 3% to 10%). The agreement between the two lidar-based tree cover data sets was high (  r 2 ≈ 0.8, RMSE ≈ 10% and bias r 2  = 0.66 ˜ 0.76, RMSE = 16% ˜ 30%, bias = 10% ˜ 24%), caused by both definitional differences and estimation errors. Because of the capability to simultaneously derive tree cover and tree height at a high accuracy, and the increasing availability of lidar data, we recommend incorporating lidar data as a complement or an alternative to ultra-fine resolution images in training Landsat-class images (e.g. Sentinel-2AB) for more precise mapping of forest extent.